AI Detector False Positives on Marketing & Freelance Copy: Every Published Number (2026)

1 What are the real false-positive rates on professional writing?

The study is Jabarian and Imas, “Artificial Writing and Automated Detection”, published as NBER Working Paper 34223 in August 2025 and also circulated as Chicago Booth’s Becker Friedman Institute Working Paper 2025-116. It is a working paper, not peer-reviewed — worth stating plainly. It remains the strongest evidence available on this question, and it is the only one that used professional prose rather than student essays.

Read the blog row again: 0.50%, 0.00% and 0.00%. If you write long-form blog content and a client tells you a detector says your work is AI, the base rate says the detector probably didn’t. Something else went wrong — and section 4 covers what that something usually is. This is a different failure mode from the one driving false positives against neurodivergent writers, where a genuine stylistic signal is being misread.

That result is consistent with what we found when we examined Pangram’s near-zero false-positive claim in detail — the low false-positive figure holds up; its false negatives are the more interesting problem.

2 Why short copy is the actual danger zone

This is the single most useful finding for commercial writers, and it inverts the usual advice. Long-form is safe. The short, formulaic, fragmentary text that fills most commercial briefs is where detectors get twitchy — because there simply isn’t enough signal in 40 words to distinguish a human from a model.

Pangram’s own published per-domain table points the same way, and the company is unusually candid about it. Its two worst non-poetry categories are the two formats SEO teams ship constantly.

Pangram goes further than its competitors and recommends against running detectors on “short bullet point lists and outlines… very short responses, and extremely formulaic text such as long lists of data, spreadsheets, template based writing, and instruction manuals.” That is a detector vendor telling clients not to screen most of the furniture on a commercial web page. If a client is running a detector over your product descriptions, the vendor’s own guidance says they shouldn’t be.

3 What the vendors claim about themselves

When GPTZero benchmarked itself against competitors on a 3,000-sample mixed-genre corpus that included news articles and blog posts, it reported a 0.24% false-positive rate for itself, 5.26% for Copyleaks and 4.79% for Originality.ai. In a separate article claiming the same dataset, GPTZero reported its own rate as 0.13%. Both figures cannot be right.

Note the Copyleaks spread: its own claim is 0.2%; its rival measured 5.26%. That is a 26-fold gap between two interested parties with no neutral referee, and no independent study has ever measured Copyleaks on non-academic text. The pattern of detectors contradicting each other is not new — we documented how much AI detectors disagree with one another across every published head-to-head.

That interview is worth reading in full alongside our own testing of whether Originality.ai is accurate or just flags human writing. The same piece contains Gillham’s stated position that his company “advise[s] against the tool being used within academia, and strongly recommend[s] against being used for disciplinary action” — while its blog calls AI detection “essential” in the classroom. His rationale for why professionals are fair game: they produce higher volume, so “the algorithm has more chances to get it right.”

4 How much income has this cost writers?

This is the largest hole in the entire topic. The harm is real and documented in individual cases; the aggregate is unmeasured. Here is everything that exists:

The Gizmodo investigation remains, more than two years later, the only substantial piece of original reporting on this beat. It documents Kimberly Gasuras, a Bucyrus, Ohio news reporter with 24 years’ experience and no AI use, flagged by Originality.ai on the WritersAccess platform, given one warning, receiving no reply to her defence, and suspended months later for “excessive use of AI.” WritersAccess did not respond to the outlet.

Gizmodo also identified the mechanism that probably causes most of the damage, and it is not detector inaccuracy. Reporter Thomas Germain ran his own article through Originality and got “70% Original / 30% AI” — a confidence score, not a proportion of the text. He spoke to multiple writers who had to argue with clients who read that number as “30% of your article was written by a machine.” A detector can be working exactly as designed and still end a contract, if the buyer misreads the output. Writers in that position need authorship evidence that actually proves something — draft history alone did not save the copywriter in the Gizmodo piece.

One more number belongs here, with a heavy caveat. A March 2026 survey by Elorites Content reports that 85% of firms and freelancers have faced a “blame game” over AI use, and 72% say client onboarding has become harder because of detector scrutiny. It states a sample of 1,165 respondents across 22 countries. It is not a projectable industry statistic: respondents were recruited from the publishing company’s own LinkedIn network and industry contacts, participation was voluntary and self-selected, and no margin of error, response rate or questionnaire was published. Elorites Content is a content-writing agency that sells human-written content — the finding directly favours its commercial position. Cite it as one agency’s customer sentiment, not as an industry measurement.

6 The numbers that are completely made up

The search results for this topic are dominated by AI-humanizer affiliate sites, which have filled the statistics vacuum with numbers they appear to have invented. We checked each of the following and could not reach a primary source for any of them.

One more circulating claim deserves a flat warning: a named freelance writer said to have lost his job over a false accusation, whose case appears in search results with no traceable reporting behind it. We could not establish that this person exists. The verified cases are the two in the Gizmodo investigation.

There is also a widely-quoted “61% false positive rate” that is real but routinely stripped of its context — it comes from a 2023 Patterns study of non-native English speakers’ TOEFL essays, and applies to that population and that format only. It says nothing about marketing copy. We cover the underlying effect in our breakdown of how detector accuracy varies by language.

7 What nobody has measured

An “every published number” page is only as honest as its account of the numbers that don’t exist. These are verified absences, each checked against academic databases, vendor disclosures, platform documentation and case law.

The Grammarly gap is the most consequential one. Gizmodo spoke to writers who were fired by platforms that required them to use Grammarly, and to detection companies who said grammar tools can trigger flags. Grammarly’s head of education disputed this on record: “There is no evidence linking AI detection flags and the use of Grammarly suggestions.” She is technically correct — because nobody has run the study. Separately, a February 2025 University of Maryland paper by Saha and Feizi found that minimally polishing human text with GPT-4o produced detection rates of 10% to 75% depending on the detector, which is the closest anyone has come to testing the editing-tool question. What that means in practice tracks closely with our findings on which AI writing markers survive editing.

Finally, the fear that sells most detector subscriptions to marketing teams is not supported by the platform it invokes. Google’s on-record statement to Gizmodo: “It’s inaccurate to say Google penalizes websites simply because they may use some AI-generated content. As we’ve clearly stated, low value content that’s created at scale to manipulate Search rankings is spam, however it is produced.” Google does not use, endorse or recommend third-party AI detectors, and has never published an AI-detection ranking signal. Originality.ai markets itself as a way to “future proof your site on Google,” and its CEO told Gizmodo that fear of de-indexing is “increasingly the No. 1 selling point for AI detectors.” The platform being invoked says the premise is wrong, and the freelancer absorbs the cost of the misunderstanding — a dynamic that also shows up in how AI detection is being used in hiring.